Abstract
Large-scale image retrieval relies on hashing to map high-dimensional data into compact, searchable binary codes. Semi-supervised hashing (SSH) methods attempt to leverage unlabeled data through pseudo-labeling; however, these approaches often suffer from accumulated bias due to label noise and optimization shortcuts. We present Generalized Debiased Semi-supervised Hashing (GDSH), a unified framework that systematically mitigates bias across the labeling, learning, and optimization stages. By employing robust label completion, sparse error modeling for denoising, and direct discrete optimization, GDSH achieves state-of-the-art performance across six major image benchmarks, including ImageNet and MS-COCO.
Introduction
In the era of big data, efficient image retrieval necessitates the use of hashing—transforming complex images into short binary strings (e.g., 11010...). To avoid the prohibitive cost of manual labeling, researchers utilize semi-supervised hashing, which generates "pseudo-labels" for unlabeled samples.
A significant challenge in this domain is the propagation of error. Initial pseudo-labels are often inaccurate, creating a "pseudo-label noise" that introduces bias. This bias is typically amplified during the learning phase and further hardened by mathematical relaxations used during optimization. Existing solutions often fail to address all these stages or are too computationally expensive for real-world scale. GDSH addresses these issues by identifying and neutralizing bias at three critical junctures: labeling, learning, and optimization.
Methodology
The GDSH framework is composed of three integrated modules designed to ensure the integrity of the generated hash codes.
1. Robust Label Completion (RLC)
Traditional label completion often relies on simple linear mappings that are sensitive to outliers. RLC instead utilizes a robust classification hyperplane governed by $\ell_{2,1}$-norm regularization. This mathematical approach encourages sparsity, preventing an image from being erroneously associated with an excessive number of categories and resulting in more reliable initial pseudo-labels.
2. Debiased Hash Learning (DHL)
The DHL module performs a two-step refinement. First, it models pseudo-label noise as a sparse error matrix, which is subtracted to recover cleaner semantic information. Second, it dynamically updates similarity relationships between samples as label quality improves. This ensures that the resulting hash codes remain semantically consistent throughout the training process.
3. Discrete Optimization
Many hashing algorithms "relax" binary constraints—treating -1 and 1 as continuous values—to simplify the mathematics. This introduces "quantization bias" when the values are eventually rounded back to binary. GDSH avoids this by using Discrete Cyclic Coordinate Descent (DCC) to optimize the binary codes directly, preserving the structural integrity of the hash space.
Experimental Evaluation
Datasets and Setup
The researchers evaluated GDSH on six widely recognized image benchmarks:
- Single-label: CALTECH-101, CIFAR-10, and ImageNet.
- Multi-label: MS-COCO, NUS-WIDE, and MIRFlickr.
These datasets provide a comprehensive testbed, ranging from simple object classification to complex scenes with multiple overlapping semantic categories.
Key Results
GDSH consistently outperformed 18 state-of-the-art hashing methods across all tested bit lengths (12, 24, 32, 48, and 64 bits).
Significant findings include:
- ImageNet Performance: On the massive ImageNet dataset, GDSH outperformed the leading semi-supervised baseline by 6.6% in mean Average Precision (mAP) for 64-bit codes.
- Multi-label Accuracy: On the NUS-WIDE dataset, GDSH showed a significant improvement of over 3.3% compared to previous state-of-the-art methods.
- Convergence and Efficiency: The training process is highly efficient, typically converging within 10 iterations while maintaining a lower memory footprint than competing denoising strategies.
Discussion and Insights
A core insight of this work is that bias is a multi-stage phenomenon. It is born during labeling, raised during learning, and hardened during optimization. By treating noise as a sparse matrix (assuming most labels are correct and only a few are erroneous), GDSH can use efficient $\ell_1$-norm optimization to filter out inaccuracies. Furthermore, the use of a sampling ratio ($\beta$) for similarity preservation allows the model to scale to massive datasets like ImageNet without a prohibitive computational burden.
Conclusion
Generalized Debiased Semi-supervised Hashing (GDSH) represents a significant step forward in large-scale image retrieval. By systematically addressing bias at every stage of the hashing pipeline, it provides a robust and scalable solution for real-world applications. Future work will explore self-adaptive techniques to automatically determine the optimal number of pseudo-labels ($k$) for each image and further optimize the algorithm for real-time streaming data environments.
Summary updated